Individuals, Institutions, and Innovation in theDebates of the French Revolution

Alexander T. J. Barrona, Jenny Huanga,d, Rebecca L. Spangb, Simon DeDeoc,d,∗

aSchool of Informatics, Computing, and Engineering, Indiana University, 919 E 10th Street, Bloomington, IN 47408, USAbDepartment of History, Indiana University, 1020 E Kirkwood Ave, Bloomington, IN 47405, USA

cDepartment of Social and Decision Sciences, Carnegie Mellon University, 5000 Forbes Avenue, BP 208, Pittsburgh, PA 15213, USAdSanta Fe Institute, 1399 Hyde Park Road, Santa Fe, NM 87501, USA

Abstract

The French Revolution brought principles of “liberty, equality, and brotherhood” to bear on the day-to-day challengesof governing what was then the largest country in Europe. Its experiments provided a model for future revolutions anddemocracies across the globe, but this first modern revolution had no model to follow. Using reconstructed transcriptsof debates held in the Revolution’s first parliament, we present a quantitative analysis of how this system managedinnovation. We use information theory to track the creation, transmission, and destruction of patterns of word-use acrossover 40,000 speeches and more than one thousand speakers. The parliament as a whole was biased toward the adoption ofnew patterns, but speakers’ individual qualities could break these overall trends. Speakers on the left innovated at higherrates while speakers on the right acted, often successfully, to preserve prior patterns. Key players such as Robespierre (onthe left) and Abbe Maury (on the right) played information-processing roles emblematic of their politics. Newly-createdorganizational functions—such as the Assembly’s President and committee chairs—had significant effects on debateoutcomes, and a distinct transition appears mid-way through the parliament when committees, external to the debateprocess, gain new powers to “propose and dispose” to the body as a whole. Taken together, these quantitative resultsalign with existing qualitative interpretations but also reveal crucial information-processing dynamics that have hithertobeen overlooked. Great orators had the public’s attention, but deputies (mostly on the political left) who mastered thecommittee system gained new powers to shape revolutionary legislation.

Keywords: cultural evolution, political science, cognitive science, digital history, rhetoric, computational social science

∗To whom correspondence should be addressed.Email address: sdedeo@andrew.cmu.edu (Simon DeDeo)

Preprint submitted to — October 20, 2017

arX

iv:1

710.

0686

7v1

[ph

ysic

s.so

c-ph

] 1

8 O

ct 2

017

The French Revolution was a turning point in Europeanhistory. Revolutionary commitments to individual libertycollided with ideals of social equality, while the rejectionof Divine-Right monarchy and the embrace of laws basedon reason opened a host of practical questions about howto govern the most populous state in Europe. The firstparliament of the Revolution, the National ConstituentAssembly (NCA), was, not surprisingly, a picture of up-heaval from its outset.

Over the course of two years, the thousand or moreindividuals in that Assembly took it upon themselves topropose and argue the previously unimaginable: the revo-cation of Old-Regime privilege and the reinvention of therelationship between individual and state. But this parlia-ment was more than a debate society for ambitious youngmen. It was also the origin of a system of rule. In the yearsthat followed, successive legislative bodies declared war onmost of Europe, dissolved the French monarchy, declareda Republic, and sentenced the former king to death—allwhile simultaneously writing constitutions and passing or-dinary legislation. Many of their procedures, and someof their personnel, were drawn from the experience of theNCA.

As a parliament, the NCA faced the problems thatcome with managing massive flows of information fromthe outside—problems that face the modern deliberativepolitical bodies that, in many cases, are the NCA’s di-rect descendants [1]. But as the first parliament theyalso shared the challenges of the knowledge-seekers of theEnlightenment and their descendants: speakers faced therisks and rewards of innovation as they introduced ideasand attempted, under unusual conditions, to achieve last-ing influence over the future arguments of their contempo-raries [2].

Key questions arise from seeing the NCA as simulta-neously a site of political and epistemic activity. On theepistemic side: How did new ideas enter that parliament-room, and how were they taken up, or discarded, by thosewho heard them? What kinds of roles did individuals playin this process? On the political side: What institutionsdid the parliament evolve to manage the onslaught of nov-elty and reaction, optimism and grievance, philosophicalargument and organizational minutiae that characterizedthe day-to-day tasks of governance and nation-building?

The digitization of historical archives allows us to an-swer these questions in a fundamentally new way. Usinglatent Dirichlet allocation [3] and new techniques in infor-mation theory drawn from the cognitive sciences [4], wetrack the emergence and persistence of word-usage pat-terns in over 40,000 speeches made in the NCA and laterreconstructed in the Archives Parliamentaires (AP) fromdetailed records kept at the time of the Revolution. Twocritical measures, novelty (how unexpected a speech’s pat-terns are, given past speeches) and transience (the extentto which those patterns fade away, or, conversely, persistin future speeches), allow us to trace both emergent po-litical institutions and the generation and propagation of

new ideas and manners of speech. Viewing the turbulentearly days of the French Revolution through the lens ofpattern creation, sharing, and destruction provides a com-plementary viewpoint to the study of specific ideas andparticular historical events. This distinction can also befound in evolutionary biology, where one can study either(on the one hand) the mechanisms of transmission andselection or (on the other) the particular phenotypes forwhich an environment selects.

We find, at high significance, a bias in favor of the prop-agation of novel patterns from prior speeches into the fu-ture: in the framework of cultural evolution, the flow ofideas through NCA is out of equilibrium and the systempreferentially selects for what violates prior expectations.This effect is driven in part by charismatic political rad-icals such as Robespierre and Petion de Villeneuve, whonot only introduced new patterns at higher rates than theirpeers, but managed to do so in a way that led others tocopy those patterns forward. By contrast, influential con-servative figures such as Abbe Maury and Cazales acted asinertial dampeners, successfully preserving prior patternsin the face of an overall revolutionary bias towards inno-vation. Conservatives “conserve”: not just by referringto past traditions, but also in the day-to-day dynamics ofdiscussion, where they act to keep conversations on track.

In parallel with these individual-level differences, ourmethods reveal a major transition in how the parliamentas a whole processed novelty, occurring about half-waythrough the parliament’s lifetime. This shift is associatedwith committees that reported to, but met outside of, theparliament itself, and that had gained new power both toraise and resolve questions. While orators on the left andright continued to confront each other in public speechesfrom the floor, it was the left that captured this new in-stitutional mechanism and used it to their own advantage.The consolidation of this structural shift was accompaniedby radicalization of the left and an accelerating flight ofconservatives from both the parliament and the countryitself.

Results

Social systems are characterized by heteroglossia: thecoexistence, sharing, and competition of different patternsof speech. Heteroglossia makes linguistics and rhetoric,both concerned with the study of the reception, influence,and propagation of language within a community [5, 6],core components in the quantitative study of culture.Tracking changes in speech patterns within a social bodyallows us to examine cultural evolution: the circulation,selection, and differential propagation of speech patternsin the group as a whole (rather than, say, tracking the ideasof a single individual). Patterns of heteroglossia demon-strate existing power relations, create new ones, and area key method for the definition of both institutions andgenres [7, 8, 9]. Our methods here quantify a key aspect

2

of cultural evolution: the extent to which one agent’s lan-guage patterns are used and copied by another [10, 11].

To study the flow of rhetorical influence and attention inthe NCA over time, we characterize how patterns of lan-guage use, uncovered by topic modeling, are propagatedfrom speech to speech. We do so using Kullback-LeiblerDivergence [12, KLD]: KLD, or “surprise”, from one pat-tern to another measures the extent to which the expec-tations of an optimal learner, trained on the first pattern,are violated by the second. Other work has demonstratedthat surprise (in the Kullback-Leibler sense of the word)is a cognitive as well as an information-theoretic quantity.It predicts, for example, what a subject will look at in adynamically-evolving visual scene [13]. KLD can be usedto map an individual’s higher-level activities as well, de-tecting, for example, biographically-significant transitionsin a subject’s intellectual life [4].

Here, we use surprise to analyze a corpus of speeches bymany different individuals. Surprise then measures boththe extent to which a particular speech deviates from thepatterns of prior speeches (novelty), and from the patternsthat will appear in the future (transience). High surprisecompared to the past indicates the topic mixture is newcompared to previous speeches, hence the term “novelty”;high surprise compared to the future indicates that laterspeeches do not retain that pattern very strongly, hencethe term “transience”.

Innovation Bias

Speeches in the NCA span a wide range in both noveltyand transience; Fig. 1 summarizes the system at the levelof individual speeches, in this case at the relatively rapidtimescale (window width, w) of seven speeches. While themajority of speeches concentrate near the symmetry line—speeches with high novelty are likely to have similarly hightransience—two results stand out. First, the scatter islarge: there are many speeches that lie far off the novelty–transience line of equality, and it is easy (for example) tofind speeches with top-quartile novelty that have bottom-quartile transience. “What is new is quickly forgotten” isa useful heuristic, but holds only in the average. Below, weconsider two potential drivers of this diversity of reception:the speaker, and the context in which the speech was made.

Second, novel speeches are unexpectedly influential. Wequantify this with “resonance,” the imbalance betweennovelty and transience (see Materials and Methods). Res-onance, the quality of at once breaking from the past andinfluencing the future, increases with novelty, as shownin the rightmost plot of Fig. 1. We refer to this positiverelationship as innovation bias: penalties to high noveltyspeeches are lower than expected in a system at equilib-rium. This bias is measured by Γ, the slope of the novelty–resonance line; positive Γ indicates innovation bias. Wefind this bias in place from the most rapid timescales (onespeech to the next, w = 1), all the way up to w ≈ 100,timescales equivalent to a day or so (see SI). This innova-tion bias lasts for at least the course of a day, as speakers

deliberately turn to new topics, but fades away on longertimescales.

Organizational roles and individual strategies

In choosing when to speak and what to say, speakers hadsome control over the relative novelty of their speeches. Aspeaker’s attitude toward the Assembly as an institution,as well as other political or philosophical commitments,would contribute to his willingness to create new word pat-terns or copy those of earlier speakers. Conversely, speak-ers had much less control over the reception of what theysaid. Idiosyncratic properties of the speaker, ranging fromthe political (e.g., party membership) and the social (e.g.,demeanor or prestige) to the rhetorical (everything fromword choice to pitch and volume of speech) would have al-tered the reception of their words, raising or lowering thetransience of the patterns they create.

Differences in strategy and reception at the individuallevel, in other words, could break the system-level trendsin favor of innovation. An unskilled but risk-tolerant dele-gate might have succeeded only rarely, so that his speeches,while novel, tended to have low, or even negative, res-onance. Conversely, the prestige or social power of aspeaker who wished to maintain a line of argument mighthave allowed him to simultaneously maintain low noveltyand high resonance, effectively keeping the conversationon track.

Individual strategies were not the only source of devi-ation from system-level innovation bias. From the firstdays of its meeting, the Assembly organized itself in sucha way as to assign explicit roles to particular speakers.These information-processing functions over-ruled an indi-vidual’s personal characteristics. An example in the NCA,and common to many later parliaments today, is the roleof president, who served as both a point of contact for theKing and an enforcer of the daily agenda [14]. The pres-ident was tasked with a code of conduct and functionalrole, and over the course of the NCA 49 delegates servedin this position.

The NCA also created another specialized entity: thecommittee. Committees, whose members were notionallyselected on the basis of relevant expertise, deliberated inprivate. They developed content outside the debate pro-cess and then presented it to the full body for public re-view. While a speaker on the floor of the Assembly mightplay to the audience in the visitor’s galleries, committeemembers addressed only each other.

Lexical markers in our data identify when a committeeproxy was speaking, and allow us to classify their speechinto two categories: “new-item” speech and “in-debate”committee speech (see SI). New-item speeches introducedofficial content to the Assembly floor, typically articles anddraft decrees to be reviewed by the legislative body, andmark transitions in attention from one topic to another.In-debate speeches occurred when a committee member orreporter engaged with other delegates following the intro-duction of the item.

3

0 2 4 6 8 10

Novelty N

0

2

4

6

8

Tra

nsi

en

ce T

Transience v. Novelty

sample x=y

counts100

101

102

7 5 3 1 1 3 5 7

Relative speech position

2

4

6

8

KL

D

Resonant sample KLDs

0 2 4 6 8 10

Novelty N

5

0

5

Reso

nan

ce R

Resonance v. Novelty

sample fit

counts100101102

Figure 1: Novelty, transience, and resonance in the French Revolution. Left: a density plot of transience vs. novelty at scale w = 7. Speechesclose to the identity (x = y) line have equal surprise from the past and to the future. Resonant speeches, whose transience is anomalously lowcompared to their novelty, break this symmetry and fall below the identity line; the middle plot shows this asymmetry for a marked sampleof high-resonance speeches. Right: resonance vs. novelty, with regression line. Although novelty is tied to transience, and therefore risk ofirrelevance, it is also directly proportional to resonance.

To understand how both individual-level strategies andsystem-imposed roles affected the production and recep-tion of new patterns, we consider the novelty and reso-nance of speeches given by those holding three distinctroles in the Assembly: the 40 most common orators, thePresident (regardless of who held the position at the time),and committee proxies. We calculate both the averagenovelty and resonance for each category (scaled by z-score), z(N ) and z(R), tracking both potentially idiosyn-cratic novelty-seeking (or avoidant) strategies, and howthey are received by the system as a whole. We do thisat scales of w from one (one speech compared to eitherthe next, or prior, speech) to 5,000 (speech compared to,roughly, three months of speeches before or after).

The system’s overall bias in favor of innovation wouldpredict that the category of speakers with the highestnovelty would also have the highest resonance. To de-termine if this expectation is validated in our sources,we compare the overall novelty-resonance relationship ofthe system (depicted by the fit line in Fig. 1) to themeasured resonance for speeches from each of the threecategories (orator, President, committee proxy). Specif-ically, for each category we report ∆z(R), defined asz(R) − E[z(R)|z(N )], the difference between the mea-sured mean resonance of speeches and the expected meanresonance under the OLS model z(R) ∼ z(N ). Whilez(R) measures the effect speakers had on later discourse,∆z(R) measures the extent to which they broke system-level trends in achieving those effects. For example, aspeaker with high novelty may have high resonance butnegative ∆z(R), indicating that his adventurousness wasrewarded less than expected. Full results for w = 36(roughly half a day) are shown in Table 1; results for nov-elty are stable on all timescales, while resonance at w = 36is strongly correlated from w = 3 to w = 100 (see SI).

Institutional roles—speaking on behalf of a committee,presiding over the Assembly—did not follow these system-level trends. High novelty and high resonance togethercharacterize committees as gatherers of new informationthat they injected into debate in ways that defined down-stream discussion. In contrast, the president’s role asagenda-enforcer led to lower than average resonance: heacted, at best, to summarize what had come before, whilehaving less influence on the patterns of speech that fol-lowed. Though he might break from conversation to fur-ther the agenda, the content he introduced to do thistended not to persist. The overall novelty-bias can notbe explained by the taking-in of new information fromcommittees: while committees did achieve above-averageresonance (z(R) greater than zero), they achieved lowerresonance than expected (∆z(R) less than zero).

Individual strategies are often as defining as top-downroles. Of the top forty speakers in the assembly, thirtyshow significant deviations from aggregate patterns in ei-ther novelty or resonance at the p < 0.05 level, withtwenty-three speakers showing deviations at p < 10−3.Speakers deviate in both directions, with some showinganomalously high tendencies to break with past patterns,and others showing similarly strong tendencies to preservethem. High-novelty speakers are overwhelmingly associ-ated with the left wing and the bourgeoisie, while all ofour right wing speakers, and the vast majority of nobility,are low-novelty.

More than half the speakers in our data pursued strate-gies that distinguished them from the aggregate; signifi-cant differences in resonance, by contrast, are less com-mon. The speakers that do show differences in reception,however, were among the key players of the Revolution.The celebrated radicals Robespierre and Petion not onlyachieved the highest average resonance in the system, but

4

Name z(N ) z(R) ∆z(R) Type

High novelty, high resonance

Jerome Petion de Villeneuve 0.10 0.28*** +0.25*** 3gMaximilien Robespierre 0.11** 0.18*** +0.14** 3gJean-Denis Lanjuinais 0.06 0.16*** +0.15*** 3gAlexandre Lameth 0.17** 0.14* +0.09 2gCharles Antoine Chasset 0.31*** 0.13* +0.04 3gCommittee (new item) 1.31*** 0.12*** -0.27*** –Philippe-Antoine Merlin 0.27*** 0.05 -0.03 3gPierre-Francois Gossin 0.65*** 0.03 -0.17** 3gJacques Francois Menou 0.40*** 0.02 -0.10 2gCommittee (in debate) 0.29*** 0.02 -0.07*** –Left Wing 0.07*** 0.02* 0.00 (g)3rd Estate 0.06*** 0.01 -0.02* –

High novelty, low resonance

Jacques Guillaume Thouret 0.16*** 0.00 -0.05 3gJac. Jo. Def. de Chapelieres 0.35*** -0.03 -0.13** 3-Francois Denis Tronchet 0.24*** -0.04 -0.11 3gArmand-Gaston Camus 0.29*** -0.04 -0.13*** 3gPresident 0.02 -0.07*** -0.08*** –Theodore Vernier 0.55*** -0.14** -0.31*** 3g

Low novelty, high resonance

Gu. Fr. Ch. Goupil-Prefelne -0.21*** 0.13** +0.20*** 3gJean-Francois Reubell -0.18*** 0.11* +0.16*** 3gLouis Simon Martineau -0.05 0.08* +0.10* 3gJacques An. Ma. de Cazales -0.44*** 0.08* +0.21*** 2dPierre Victor Malouet -0.27*** 0.08* +0.16*** 3dJean-Sifrein Maury -0.46*** 0.07 +0.20*** 1dPierre-Louis Prieur -0.27*** 0.05 +0.13** 3g1st & 2nd Estates -0.10*** 0.03*** +0.05*** –Jean-Fr. Gaultier de Biauzat -0.13** 0.03 +0.06 3gRight Wing -0.32*** 0.03* +0.10*** (d)

Low novelty, low resonance

Dominique Garat -0.13** 0.00 +0.04 3gAntoine Ch. Ga. Folleville -0.44*** -0.01 +0.12* 2dLo.-Mi. Le Pel. de Saint-Fargeau -0.20*** -0.01 +0.05 2gFr. Do. de Reynaud Montlosier -0.61*** -0.02 +0.17* 2dPierre-Louis Roederer -0.10* -0.03 0.0 3gLouis Foucauld de Lardimalie -0.53*** -0.05 +0.11 2dCharles Malo Lameth -0.15*** -0.06 -0.02 2gPierre Franois Bouche -0.09** -0.10* -0.07* 3gAntoine Barnave -0.04 -0.12* -0.11* 3g

Table 1: Mean novelty and resonance per speaker at scale 36, forrole and type (in bold) and the top forty orators. z(N ): noveltycompared to system average; z(R): resonance compared to systemaverage; ∆z(R): resonance relative to predicted resonance given nov-elty. “Type” codes for estate (3: bourgeoisie; 2: nobility; 1: clergy)and political affiliation (g: gauche, left-; d: droit, right-wing).

significantly higher resonance than even that due to thesystem-wide novelty bias (positive ∆z(R)). Conversely,speakers such as Armand-Gaston Camus and TheodoreVernier, called on primarily for their specialized knowledgein canon law and taxation, respectively, show high-novelty,but low resonance: they presented information, but eitherlacked the lasting influence of more prominent speakers, orwere able to settle questions so conclusively that the roommoved on. Finally, prominent political conservatives suchas Jean-Sifrein Maury and Jacques de Cazales appear inthe low-novelty, high-resonance quadrant. They were ableto break the system-level novelty bias, and are notable notonly for keeping the conversation on track (low novelty),but for being able to influence others to do the same (highresonance). In this, Maury and Cazales are characteristic

of the right-wing overall: while the novelty-biased left wascomposed of both high and low resonance speakers, theright-wing was able to achieve system influence (positivez(R) and ∆z(R)) despite their anomalously low novelty.

The Emergence and Evolution of the Committee

Committees were a key innovation of the NCA, and al-lowed the system to manage large amounts of informationwithout overwhelming the discussions in the parliament-room itself. They did not, however, appear overnight.While the previous section establishes their unusual func-tional role, the comprehensive coverage of the AP makesit possible to study how their role emerged. In this sec-tion, we show how returns to novelty, Γ, are modulated bycommittee roles over time. We fit, separately, two termsthat quantify the additional boost (or decrement) to thenovelty-resonance relationship when speeches either intro-duce new committee items (Γn), or advocate on behalf ofa committee during debate (Γd). A speech of novelty N ,for example, achieves on average a resonance R equal to(Γ + Γn)(N − N0) when made by a committee memberintroducing a new item, compared to Γ(N −N0) when thespeaker acts on his own behalf.

We look for discrete shifts in committee function, doingchange-point detection with a maximum-likelihood modelof the novelty-resonance relationship where Γ, Γn, and Γd

are allowed to vary in time. Following Ref. [4], we con-sider a two-epoch model, where all three quantities arefixed to constant values in each epoch, with a single dis-crete change at a particular time-point whose position isa free parameter. The two-epoch model is preferred to asingle-epoch model, as well as to a linear (secular shift)model under AIC. We find strong evidence for a change-point in the nature of committee functions occurring atthe end of 1790; the modal best-fit date across all scales is31 October 1790. Allowing the intercepts of the new-itemand in-debate speeches, as well as their slopes, to varyproduces nearly identical results (see SI). Fig. 2 displaysthe novelty-transience relationships for the two epochs,demonstrating the emergence of a new role for committeespeech over time. In the first epoch, the resonance of newitems was indistinguishable from speech of similar noveltyin the system as a whole. However, speech received a res-onance bonus when delivered by a committee member aspart of a debate. In other words, committee representa-tives injected new information in a fashion similar to otherdelegates (new-item speech), but had privileged abilities inguiding the subsequent debate (in-debate speech).

The pattern is different in the second epoch, where new-item speech has anomalously low resonance at high nov-elty. Inspection of the speeches themselves suggests thatthis high-novelty/low-resonance tail is associated not with(as in the case of individual speakers) failure to alter thecourse of debate, but rather with the increasing powercommittees had to “propose and dispose”: once the com-mittee presented their findings to the parliament, theywere increasingly accepted with minimal discussion. In

5

100 101 102 103

Scale

0.4

0.2

0.0

0.2

0.4

Γn

&Γd

First epoch

in-debate

new-item

100 101 102 103

Scale

scale=27

Second epoch

4 2 0 2 4 6

z(N)

5

0

5

z(R

)

scale=27

non-committee

in-debate committee

4 2 0 2 4 6

z(N)

scale=27

new-item committee

Figure 2: Information-processing functions of NCA committees before (1st column) and after (2nd column) the late-1790 change-point. Toprow: the shift in the novelty-resonance relationship for new-item and in-debate committee speech, with 99% confidence intervals. Bottomrow: scatter plots and fit lines at scale 27 for these speech types, compared to all other speeches. The “undebated tail” is seen in the secondepoch, as committees gain new powers to “propose and dispose”.

some cases, committee reports are noted as “passed with-out debate”; for a sample of these cases, we find the meanz(N ) equal to 1.57±0.16 and z(R) equal to 0.14±0.13, i.e.,committee speeches with these hand-annotated outcomesfollow a similarly high-novelty, low-resonance pattern toother new-item speeches, further validating the interpre-tation of the “undebated tail” as a signal of emergent com-mittee power. This power can also be seen when substan-tive debate on committee matters does occur: “in debate”speeches made by committee members retain a privilegedrole in fixing and propagating the patterns that define thatdebate.

Discussion

The turbulent months at the beginning of the FrenchRevolution led to a durable transformation in the very na-ture of European government [15]. Our study considersthe elite debates at the center of these events, focusing onpattern transmission rather than the differential propaga-tion of the ideas these patterns may communicate. Thisprovides a new window into qualitative analyses, whichusually focus on analyzing the logic of particular ideas andarguments over time [16, 17].

Our analysis reveals clear differences between how theleft- and right-wing figures created and transmitted pat-terns of language use. Conservatives may indeed “standathwart history, yelling Stop” [18]; our results show thathere they did so less by redirecting conversation from one

set of patterns to another than by maintaining the already-established patterns of the conversations in which theyparticipated. Indeed, the spatial metaphor of “left vs.right” is itself misleading: from the point of view of thedebates themselves, the right wing acts as an inertial cen-ter, holding off conversation drift, while left-wing speak-ers produce a wide spectrum of innovations on a muchlarger periphery, only some of which survive. These rolesbecome visible without reference to the ideas that arepropagated: Robespierre emerges as a high-novelty, high-resonance speaker even before we consider the content ofthe speeches that made him an icon of revolutionary poli-tics.

We also find an emergent distinction between oratorsand institution players. Talented speakers like Robe-spierre, Petion, and Maury could achieve power throughrhetoric and debate, appealing to the crowds in the gal-leries as much as their colleagues. Yet not every ambitiousdelegate could take this road. Quite apart from a dele-gate’s rhetorical capacities or political stature, the phys-ical venue itself was demanding: loud and (literally) res-onant voices were required to be heard. While the NCAprovided a place for rhetorical masters to thrive, in otherwords, it also ended up producing a second class of del-egate who concentrated on committee work. Deliberat-ing outside the parliament-room, committees contributedspecialized knowledge and abilities and became indispens-able workhorses for the emerging government [14]. Theinformation-processing functions that committees took on

6

distinguish them clearly from other forms of speech, whileFig. 2 shows how these functions emerged over time; oncein place, committees are both strong sources of new pat-terns of speech and new sources of extra-rhetorical power.Both individuals and institutions, in other words, mat-tered, playing distinct roles in the development of the par-liament over time.

Many delegates welcomed these extra-parliamentaryfunctions: committees, wrote the assembly memberJacques Dinochau, “regulate the order of debate, classifyquestions, and maintain a continuity of principles, thuspreventing an incoherence which might otherwise havemenaced our decrees” [14]. Yet the private nature of thesecommittees was in dramatic contrast to the public debatesin the hall itself; their increasing influence a testimony tothe emergent distinction between the spectacle of democ-racy and the actual ways functioned in a body and politytoo large for direct participation. The dramatic and earlyappearance of a specialized information-processing role forcommittees provides a new view on their place in moderndemocracies, where they emerge endogenously [19] to serveas essential information-management systems [20].

The power of these committees continued to grow.While early committees were devoted to technical mat-ters such as monetary and fiscal policy, they were alsoinstrumental at key moments such as the dismantlementof feudal and religious privileges. When the revolution col-lapsed into chaos in 1793, it was committees, such as thefamous “Committee of Public Safety” with Robespierre atits head, that effectively replaced the republican govern-ment and executed thousands of “enemies of the revolu-tion” including many of the speakers in Table 1.

Conclusion

The history of human culture is more than just the riseand fall of particular ideas. It is also the emergence of newinformation-processing mechanisms and media, and rolesthat individuals and institutions could play in creating andselectively propagating these ideas through time. In thelanguage of biological evolution, we must understand notonly the particular characteristics for which an environ-ment selects, but the strength of that selection over time,and the shifting and heterogenous nature of the transmis-sion mechanisms.

By quantifying the flow of patterns between speakers,we can see revolutionary debates as not just a battle ofideas, but as struggle for the direction of the conversationitself. At the earliest stages of the revolution, politicalplayers adopted not just different ideological positions, butdifferent roles in the propagation of patterns, with varyinglevels of innovation and fidelity to the past. And together,both cooperatively and agonistically, they invented newmechanisms for the collective management of information.

Materials and Methods

The AP is the definitive source for parliamentary tran-scripts of the Revolution, reconstructed from primarysources including transcripts, minutes, and newspaper re-ports. The French Revolution Digital Archive (https://frda.stanford.edu) is a digital version, with full-textspeeches and TEI markup from the beginning of the NCAon 9 July 1789 through its end on 30 September 1791.After stop-word removal, each speech is represented asa count vector over vocabulary of 10, 000 most commonwords. The resulting corpus contains 4, 765, 773 wordsin 44, 913 speeches. Each speech is matched to a namedspeaker, and may be tagged with a “role”: in some cases,a speech made by a delegate acting as the assembly pres-ident; in other cases, on behalf of a committee, either tointroduce a report, or to comment on it. See Supplemen-tary Information (SI; linked as ancillary file on the arXiv)for details on corpus preparation. All of the delegates tothe Assembly, and all of the speakers in our data, weremale.

We use LDA [3] to model speeches as probability distri-butions over K = 100 topics; the jth speech in the data is

s(j)i , 1 ≤ i ≤ K.

Novelty, Transience, Resonance

Novelty at the smallest timescales (a speech comparedto the one just previous) is measured by the KLD of s(j)

relative to s(j−1)

KLD(s(j)|s(j−1)

)=

K∑i=1

s(j)i log2

(s(j)i

s(j−1)i

). (1)

Averaging this measure further backwards in the debateallows us to consider longer-timescale trends, beyond theback-and-forth of a single exchange, to study the extentto which the current speaker has taken the debate in anunexpected direction given, say, the last twenty speeches.We refer to this quantity as novelty N at time j on scalew,

Nw (j) =1

w

w∑d=1

KLD(s(j)|s(j−d)

), (2)

illustrated in Fig. 3.

Any speech can break abruptly from its past—but thenew patterns it introduces may not persist. Consider aninterjection that other speakers ignore in order to return tothe matter at hand. That interjection would be surprisinggiven its past, but equally surprising given its future, sym-metric surprise under time reversal, because all the new in-formation is lost, or transient. In contrast, a rhetoricallyeffective interjection might shift the conversation to a newissue with differing vocabulary. This shift would appearas a surprise asymmetry about the interjection. We define

7

......j-1j-dj-w j

......j+1 j+wj+d

centerspeech

time

speeches

⎬ ⎫⎭ ⎬ ⎫⎭

Novelty Transience

Figure 3: Diagram of novelty and transience. Each speech is rep-resented by an LDA topic mixture. The Kullback-Leibler Diver-gence KLD

(s(j)|s(j−d)

)measures surprise of speech at time j given

a speech d steps away. Novelty at scale w is the mean surprise of acentral speech given speeches in a past window of w steps. Transienceis novelty under time reversal, measuring the mean surprise from thefuture. A rhetorically effective speech shifts current conversation toa new issue with a differing vocabulary, manifesting itself as a sur-prise asymmetry where novelty exceeds transience. We define thispositive asymmetry as resonance (Eq. 3): the ability to break fromthe past and influence the future. An illustration of this asymmetryis shown in the middle panel of Figure 1.

this asymmetry as “resonance”, R:

Rw (j) =1

w

w∑d=1

[KLD

(s(j)|s(j−d)

)−KLD

(s(j)|s(j+d)

)]≡ Nw (j)− Tw (j) (3)

Resonance is novelty minus transience, T , where the lat-ter is novelty in Eq. 2 under time reversal. Novel speecheswhich also influence future discourse are pivot points inconversation. For a system in equilibrium, deviationsare time-symmetric in the expectation—i.e., the measurednovelty and transience at any point in time are, on average,equal. For a non-equilibrium system significant deviationsmay appear both for individual samples and for the over-all average. Novelty’s effectiveness, Γ, is the rate at which

resonance increases with novelty, dE[R|N ]dN ; approximated

here using a linear model.

Acknowledgments

J.H. acknowledges a Research Experience for Undergrad-uates National Science Foundation Grant #ACI-1358567at the Santa Fe Institute.

References

[1] B. D. Jones, F. R. Baumgartner, The politics of attention: Howgovernment prioritizes problems, University of Chicago Press,Chicago, IL, USA, 2005.

[2] J. G. Foster, A. Rzhetsky, J. A. Evans, Tradition and innovationin scientists’ research strategies, American Sociological Review80 (5) (2015) 875–908.

[3] D. M. Blei, A. Y. Ng, M. I. Jordan, Latent Dirichlet allocation,Journal of Machine Learning Research (JMLR) 3 (2003) 993–1022.

[4] J. Murdock, C. Allen, S. DeDeo, Exploration and exploitationof Victorian science in Darwin’s reading notebooks, Cognition159 (2017) 117–126.

[5] W. L. Benoit, M. J. Smythe, Rhetorical theory as message re-ception: A cognitive response approach to rhetorical theory andcriticism, Communication Studies 54 (1) (2003) 96–114.

[6] R. A. Harris, Reception studies in the rhetoric of science, Tech-nical Communication Quarterly 14 (3) (2005) 249–255.

[7] M. Foucault, Power/Knowledge: Selected interviews and otherwritings, 1972–1977, Pantheon, New York, NY, USA, 1980.

[8] C. Danescu-Niculescu-Mizil, L. Lee, B. Pang, J. Kleinberg,Echoes of power: Language effects and power differences in so-cial interaction, in: Proceedings of the 21st international con-ference on World Wide Web, ACM, 2012, pp. 699–708.

[9] S. Klingenstein, T. Hitchcock, S. DeDeo, The civilizing processin London’s Old Bailey, Proceedings of the National Academyof Sciences 111 (26) (2014) 9419–9424.

[10] M. M. Bakhtin, The dialogic imagination: Four essays, Univer-sity of Texas Press, Austin, TX, USA, 2010.

[11] R. A. Blythe, W. Croft, S-curves and the mechanisms of prop-agation in language change, Language 88 (2) (2012) 269–304.

[12] S. Kullback, R. Leibler, On information and sufficiency, Annalsof Mathematical Statistics 22 (1) (1951) 79–86.

[13] L. Itti, P. Baldi, Bayesian surprise attracts human attention,Vision Research 49 (10) (2009) 1295–1306.

[14] T. Tackett, Becoming a Revolutionary, Princeton UniversityPress, Princeton, NJ, USA, 2014.

[15] W. H. Sewell, Historical events as transformations of structures:Inventing revolution at the Bastille, Theory and society 25 (6)(1996) 841–881.

[16] K. M. Baker, Inventing the French Revolution: essays on Frenchpolitical culture in the eighteenth century, Cambridge Univer-sity Press, Cambridge, UK, 1990.

[17] K. M. Baker, D. Edelstein, Scripting revolution: a historical ap-proach to the comparative study of revolutions, Stanford Uni-versity Press, Stanford, CA, USA, 2015.

[18] W. F. Buckley Jr., Our mission statement, National Review 1.[19] D. P. Baron, Legislative organization with informational com-

mittees, American Journal of Political Science 44 (3) (2000)485–505.

[20] K. A. Shepsle, B. R. Weingast, Positive theories of Congres-sional institutions, Legislative Studies Quarterly 19 (2) (1994)149–179.

8